How come AI hasn't advanced very much in the last decade?

[Deleted member 22585]

Basically you don't want games to have good AI because it wouldn't be any fun to play against.

Click to shrink...

I read this opinion often and am always confused.
I don't automatically link good AI to higher difficulty for the player.
For me, better AI means that enemies react more dynamic, have more interactions with the environments, more complex behaviour and patterns etc. There's a lot of possibilities to make combat more engaging without making the enemies more difficult, it can be balanced.
Also,I think of AI in the context of open world games and our current AI behaviour is very primitive.

 

[DiipuSurotu]

Good AI isn't necessarily something that would apply to enemies only. What about CPU-controller characters, etc.

Remember how shit the AI was in Secret of Mana back in the day?

 

[DarthBuzzard]

Once more developers jump on board with VR, you'll see plenty of AI advances.

The best way to improve AI is to make them more reactive, more intelligent without hindering the player. Make friendly AI smarter, make them more reactive to the player.

 

[collige]

It was Dota 2 best players in the world vs Elon Musk's "Open AI", and the players got stomped into the ground.

Elon Musk's 'Dota 2' Experiment is Disrupting Esports in a Big Way - No Playing Field

Elon Musk's artificial intelligence research company, OpenAI, is developing a self-learning bot for one of the most complex esports titles: 'Dota 2.' It has ...

www.youtube.com

People always think that advanced AI costs a ton of money or is too difficult to program but from most of the developer comments Ive seen they purposely dont pursue advanced AI because video games are still just games and games are supposed to be fun and be beaten.

Click to shrink...

OpenAI did cost a ton of money though. The example from that video is an extremely limited use case and it took them several more years of R&D to have it beat pros in a real game.

Getting enemies to completely wreck a player is trivial because you can have them just hack. Getting AI to actually behave like a person is much much harder.

 

[Carn]

I remember with one of the Civilization games, the developers were really proud at how the AI acted more like a human player and players thought the game's AI was broken when other countries would break treaties and backstab them for personal gain. Players don't actually want advanced AI enemies - they want smoke & mirrors that makes it seem like the enemies are smart while still allowing the player to easily beat them (like in F.E.A.R).

Click to shrink...

Yep, this is mostly it. I guess people would like 'more creative / clever' AI but still be able to beat them; which is kinda the opposite of each other. A lot of games are puzzles if you look at them; you want to analyse the enemies patterns, know their behaviours, and make use of them. I do think that more impressive NPC AI might create more living worlds.

 

[Thera]

Is it because it's brutal/expensive to program AI or is it that it's not as easy to market/sell AI rather than flashy graphics?

Click to shrink...

Both.
For exemple, we know Kojima like AI in his game. But Death Stranding has one of the moste stupid AI I exeprienced in a long time.

It's not something that comes off as exciting from fixed screenshots and trailers, so devs rarely bother.

Click to shrink...

Devs do what studios are telling them to do, they don't choose to make poor AI.

Smart AI frustrates players

Click to shrink...

I often read that but that's BS :

• Which games sold terribly or have a terrible reputation because of that?
• There is a thing called difficulty settings, sliders, tweaks, ...

And why people think good AI mean godlike AI ? Good IA is believable behavior, not wall hack and auto aim.

 

[werezompire]

Good AI isn't necessarily something that would apply to enemies only. What about CPU-controller characters, etc.

Remember how shit the AI was in Secret of Mana back in the day?

Click to shrink...

That's not entirely without its perils though. For example, a common complaint about FF13 was that the AI was so good that you could put the game on auto and it would play better than some players.

 

[Deleted member 49611]

To be fair even natural intelligence hasn't advanced in the last decade

Click to shrink...

lol if anything it feels like it has degraded.

Jaguar CPUs most likely.

Click to shrink...

love your avatar btw. WATP.

 

[Kill3r7]

Publishers haven't figured out how to sell good AI. Also, really smart AI hampers the power fantasy (aka fun) in certain genres.

 

[headspawn]

I'm not sure how much people actually want AI to be advanced when every other day you see an outpouring of the need for easier modes; better AI would obliterate you and I don't think it'd be that hard for a developer to train their AI with machine learning to develop sharper tactics to stomp you, but then MOST wouldn't find that fun. So is it worth the investment for something that would be more towards hardcore players and detrimental for others? Hell people cry about Souls games, and those games have braindead AI, could you imagine if the enemies worked together and employed tactics? For me that'd be awesome, I'm not sure an actual hard Souls game would sell well though.

 

[Old Gimlet Eyes]

It's never happening. Play online multiplayer games to get away from the patterns and predictability of garbage single player games.

 

[mute]

The solution is make your game online if you need to solve a complex AI problem. See Gran Turismo Sport.

 

[Apathy]

It's not that they can't do it (programmers are smart people, they have been able to make better ai for a long time), it's that truly good ai can turn off player base.

Video Games are a power fantasy where you are amazing and can best anything thrown at you. Good ai that actually challenges that would hurt the ego of way too many players.

 

[Hilarious_Bearious]

I think when people say they want better AI, they really mean they want more elaorate scripted behaviour. It's nice to see NPCs do lots of little things make them appear more human, but we don't want every enemy to be Machiavelli.

 

[Ewaan]

love your avatar btw. WATP.

Click to shrink...

Haha thanks mate!

 

[4 Get!]

Everyone here is just thinking in terms of combat/enemy A.I. which is why this thread's responses are mostly saying 'this is a bad idea' but what about Town A.I. in an open world RPG? They could actually be doing things other than following a pathing placed for them and it could lead to more randomized events.

Edit: Also I find it weird that people keep mentioning the power fantasy yet we constantly applaud and hype up extremely difficult online and offline games such as the Souls series. There's a disconnect happening here.

 

[Starlatine]

For me, better AI means that enemies react more dynamic, have more interactions with the environments, more complex behaviour and patterns etc. There's a lot of possibilities to make combat more engaging without making the enemies more difficult, it can be balanced.

Click to shrink...

All of this things would make combat less predictable and thus more challenging. You can't have better AI without increasing the difficulty simply because the enemy isnt as dumb and exploitable as the average AI anymore. People wont be able to exploit patterns anymore, and they will get frustrated. And this is on an industry that already codes several handicaps in favour of the player (enemies missing first shot, hit rates higher than the average for the player and lower for the enemy despite the hit percentage says etcetera)

Someone asked for examples of it, and Alien Isolation while overall praised for its tense atmosphere and great use of the Alien, most people complain about it in the harder difficulties thinking its cheating to get the player (it isn't). People get mental if you take the flamethrower away cause its the guaranteed, 100% cheese against the AI and without it the alien being too smart becomes "annoying". If they made the AI smarter thus the flamethrower less effective, i don't think most would have even finished it (not that i blame them Isolation is loooong)

 

[TheClaw7667]

It's not that they can't do it (programmers are smart people, they have been able to make better ai for a long time), it's that truly good ai can turn off player base.

Video Games are a power fantasy where you are amazing and can best anything thrown at you. Good ai that actually challenges that would hurt the ego of way too many players.

Click to shrink...

People always say this but I really don't think people mean they want AI that makes games more difficult when saying they want better AI. We already have games with basic AI and they are some of the hardest games in the medium. They want more believable AI.

There is nothing stopping programmers from making AI more believable without making the games any harder while keeping the power fantasy most games want to give players. People bring up FEAR as an example of good AI because it's AI was believable, not because the AI was so hard that it couldn't be beaten. With more believable AI the power fantasy isn't diminished instead it's increased so I doubt the reason we haven't seen much of an improvement in believable AI is because of the lack of power fantasy affecting players' ego.

I think the reason we haven't seen much of an improvement in believable AI is that it's damn hard to do. Any AI programmer can make the AI so good it can make a game unenjoyable, but to make AI fun and believable is a lot more challenging while also taking up resources that could be used for more marketable improvements. For now, most AI in games do a good enough job for their gameplay loop to work and be fun. Hopefully, we'll see some new AI techniques being implemented in next-gen games with improved CPU's in the new consoles.

 

[Starlatine]

There is nothing stopping programmers from making AI more believable without making the games any harder while keeping the power fantasy most games want to give players. People bring up FEAR as an example of good AI because it's AI was believable, not because the AI was so hard that it couldn't be beaten. With more believable AI the power fantasy isn't diminished instead it's increased so I doubt the reason we haven't seen much of an improvement in believable AI is because of the lack of power fantasy affecting players' ego.

Click to shrink...

Enemies that seek cover and run away from grenades and assorted explosives instead of directly into them are by default "harder". F.E.A.R. is indeed slightly more "difficult" than the average shooter thanks to an slightly more intelligent AI. You can't separate one from other. A smarter AI is harder to kill. It's pure logic, i don't know why so many people are arguing against this.

Now if theres anything wrong with that or not is up to anyones taste

 

[laxu]

I read this opinion often and am always confused.
I don't automatically link good AI to higher difficulty for the player.
For me, better AI means that enemies react more dynamic, have more interactions with the environments, more complex behaviour and patterns etc. There's a lot of possibilities to make combat more engaging without making the enemies more difficult, it can be balanced.
Also,I think of AI in the context of open world games and our current AI behaviour is very primitive.

Click to shrink...

Exactly. Better AI does not have to mean specifically combat AI at all.

If you play something like RDR2 you are at first impressed by all the life and unique situations that happen in the world..until the systems fall apart. Then you just see the overly aggressive lawmen and bounty hunters who will try to gun you down over a $5 bounty. You see the dialog trees fall down by not being able to craft conversations between Arthur and NPCs that make sense.

These are the kind of things I hope become better in next gen games, whether it's by making NPCs into more complex state machines or some actual AI driven behavior. Whatever works to make "random NPC #123" become someone who I could consider being an actual character in the world rather than a prop. I want to see a "RDR 3" where the AI does not jump to conclusions right away, one where I can try to reason with them before the guns come out and AI that doesn't immediately know what's up. Nearly all games still have exactly two states for characters - non-aggressive and "I'm gonna kill you". They don't go "whoah where did you come from, who the hell are you?" when you get spotted, they go into "I don't know you so I must kill you" which isn't exactly human behavior.

TLOU2 is a good modern example of employing the Fear method of making the enemies come alive. They each have names, they will call for their friends and go investigate if they think they are missing. Of course the game knows they were killed and just plays a preset scenario but to the player it looks like the enemies have a semblance of intelligence because the enemies are talking all the time. This also gives the player cues about what is going on, reinforcing the gameplay loop.

 

[Bulebule]

Advanced/somewhat believable AI in terms of combat/action/stealth is a huge turnoff for mainstream audience if there's even a subtle hint of making it tougher. Less mainstream appeal -> less money.

Also, good AI to make world feel alive is not going to be a priority either because graphics sells more. Simpler the system the less it costs and even less people care about it.

 

[Ryuelli]

I think when people say they want better AI, they really mean they want more elaorate scripted behaviour. It's nice to see NPCs do lots of little things make them appear more human, but we don't want every enemy to be Machiavelli.

Click to shrink...

This is what I want. As someone who prefers to play games on easy because I look at them mostly as a form of relaxation, I don't want "harder" AI, I just want AI that feels more alive.

 

[Grue]

I read this opinion often and am always confused.
I don't automatically link good AI to higher difficulty for the player.
For me, better AI means that enemies react more dynamic, have more interactions with the environments, more complex behaviour and patterns etc. There's a lot of possibilities to make combat more engaging without making the enemies more difficult, it can be balanced.
Also,I think of AI in the context of open world games and our current AI behaviour is very primitive.

Click to shrink...

I think when people say they want better AI, they really mean they want more elaorate scripted behaviour. It's nice to see NPCs do lots of little things make them appear more human, but we don't want every enemy to be Machiavelli.

Click to shrink...

Agree with both of these points.

In the context I think the question was intended, 'improved AI' doesn't mean either 'more like an actual human' or 'harder'.

It can mean more complex and leading to more interesting results, and more immersive, interactive or downright fun videogames. Those things have nothing to do with acting more like an actual human, or challenge.

To the question itself:

I don't have expertise so I'm prepared to be shot down here, but I suspect part of the reason is that AI relies more on software than hardware. Even with minimal effort, graphics advance naturally in line with hardware. AI doesn't, and requires much more effort to be meaningful and interesting within your specific game's environments.

 

[Vinx]

OpenAI did cost a ton of money though. The example from that video is an extremely limited use case and it took them several more years of R&D to have it beat pros in a real game.

Getting enemies to completely wreck a player is trivial because you can have them just hack. Getting AI to actually behave like a person is much much harder.

Click to shrink...

The Dota 2 OpenAI bots went from reliably beating pro players in 1 versus 1 matches to reliably beating them in 5 versus 5 matches in 2 years.

Also, how does a person behave? How are we supposed to program AI to behave like people when we are still guessing at what human behavior even is?

 

[Tygre]

A bunch of reasons, but the most important ones:

There's no evidence that games having better AI leads to more sales, or even to 'better' games. Most if not all of the most profitable and popular games in history have either no AI, or extremely basic AI. Why spend the money and effort on a feature for an extremely nebulous and doubtful return?

AI is hard. What most people think of as "Good AI" is actually good level design, mixed with good pathing, mixed with good combat barks and animations. People bring up FEAR a lot, but as the developers and designers of that game have explained many times, it was 100% creative smoke and mirrors.

AI does not scale simply with processing power. You want a game to run at higher resolution / higher framerate / with more particle effects on screen at once? More processing power can make that happen without much if any additional work. You can't simply throw more processing power at the problem of "AI" and have it magically get better.

 

[bionic77]

Games from the big companies are so gigantic now.

Its not like in the beginning when you had a dozen people working on a team and the financial hit from trying something new or crazy was worth it.

I would imagine if you are making a game that takes 5 years everything has to be set pretty early on (i.e. graphics engine, AI system, etc) and then you fill in the blanks as you try to finish it. It would probably be very difficult and risky for a developer to spend significant resources developing an entirely new AI routine and making the game around it.

 

[The Iron Banana]

Because games have such simple AI it's become a joke. Too tough though and people complain. It's expensive to implement and the payoffs just don't seem to be worth it.

 

[Ghostwalker]

Because games have such simple AI it's become a joke. Too tough though and people complain. It's expensive to implement and the payoffs just don't seem to be worth it.

Click to shrink...

People keep saying AI is expensive to implement is this based on any evidence or just an assumption?

As all the games with really interesting AIs for the most part have not been big-budget AAA games.

 

[DieH@rd]

The primary purpose of the enemy AI is not to be realistic, but to be fun. Fun challenges are powered by semi-predictable enemies that often telegraph their moves so that player can react accordingly.

Normal enemy NPC lives on screen for few seconds, that's not much time for them to showcase incredible AI.

And yes, hardware was also a limit. We'll see what Zen2 will be able to deliver soon. :)

 

[Deleted member 20852]

I think it's largely a design issue rather than a technical one. What constitutes for "good AI" largely depends on the game and what types of experiences it tries to create. Creating a perfect AI that always beats the human player is in many ways easier than creating an AI the human (let alone a bunch of people with different skillsets) enjoys playing against.

 

[Tobor]

I read this opinion often and am always confused.
I don't automatically link good AI to higher difficulty for the player.
For me, better AI means that enemies react more dynamic, have more interactions with the environments, more complex behaviour and patterns etc. There's a lot of possibilities to make combat more engaging without making the enemies more difficult, it can be balanced.
Also,I think of AI in the context of open world games and our current AI behaviour is very primitive.

Click to shrink...

Everything you just said literally translates to "more difficult".

The dirty secret is that when playing against AI most of the audience just wants punching bags we can mow down.

Why would developers spend massive amounts of money and time developing more complex AI when a huge part of the audience will turn it down?

 

[smash_robot]

I never bought into the AI is CPU intensive. It seems that game AI is just a state machine with some behavioural RNG - the reason it hasn't improved is that it's hard to implement, not that it's resource intensive.

I would hope that over the recent years that ML is being used to at least generate the state machines, but who knows.

 

[Lunchbox]

No money in it

 

[jotun?]

Smart, natural-feeling intelligence just isn't as fun to play against in most games, and would be more difficult to design around. It's easier for design to have AI that's predictable, and that also lets players learn and exploit their patterns as a way of improving at the game.

 

[monketron]

I generally avoid PvP because I just don't have the time or inclination to get good enough to compete with people who are clearly better at the game than I am. Imagine a PvE environment where it's as sweaty as PvP because the enemies are 'smart'. Urgh no.

 

[Mechaplum]

I think there is space for enemies to behave a bit more reactively, but proper emergent AI won't be very fun for most of the genres. Games are recreational, so the focus is mostly to reward players with pattern recognition and feedback loop learning tasks.

AI really means Adequate Intelligence, even in non-games implementation.

 

[Fawz]

Bad CPUs, no viable markatability, less tangible as a concept and the popular genre of games shifting away from the focused experiences that require good AI

As games get more complex it's also a lot more complicated to make it work, often it's much more viable to simply focus on making AI look like they're smart with animations and dynamic interactions over actually making them smart from a gameplay perspective

 

[waugh]

There's a misconception that the majority of gamers want harder games.

Click to shrink...

Exactly. See what happens whenever a new From Software game comes out. There will be several threads and journalist think pieces about why From games need an easy mode. People don't like the idea that they can't play something. Especially when the game has an appealing art style. See Cuphead.

 

[collige]

The Dota 2 OpenAI bots went from reliably beating pro players in 1 versus 1 matches to reliably beating them in 5 versus 5 matches in 2 years.

Click to shrink...

That was my point. A completely separate company spending 2 years with some of the best AI researchers (and the hardware required to run all the training) is a serious undertaking. OpenAI beating pros is a massive achievement.

Also, how does a person behave? How are we supposed to program AI to behave like people when we are still guessing at what human behavior even is?

Click to shrink...

We don't need to guess at what human behavior is, that's how machine learning works. The Dota 2 example isn't 100% like that since it trained against itself rather than against real people for the most part, but human behavior in the context of a game can be quantified by recording human inputs. I don't know of any games besides VF4 that have machine learning based bots though because that would be a ton of effort for a feature that most players don't give a shit about.

 

[Owlet]

I think a good example of an advance an AI that didn't make the game too punishing to be fun was Alien Isolation.

There's a couple videos that break to the AI system the alien uses that are really neat. I forget who made them, I'll have to find them.

 

[808s & Villainy]

The biggest difference between the AI in FEAR in the AI in modern games is that the former is propped to an unreasonably high, unattainable standard. You can read their white paper on how they designed their AI.

https://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf

Nothing about the design is fundamentally different from how AI is designed now. Watch various GDC talks, read Gamasutra articles, and watch Youtube BTS vids on AI design. In fact, most modern day AI does a lot of the same and even takes it a step further





And this isn't taking into account numerous games have various AI systems for things like crowd behavior, animal behavior, etc that all interact with one another in much much open environments

 

[Starlatine]

I think a good example of an advance an AI that didn't make the game too punishing to be fun was Alien Isolation.

Click to shrink...

i dont think you heard many impressions about the game
cause i sure heard lots of people saying exactly that the Alien was too punishing to be fun

complaints about it being an instakill, complaints about it "cheating" to find you, pick your poison

The biggest difference between the AI in FEAR in the AI in modern games is that the former is propped to an unreasonably high, unattainable standard. You can read their white paper on how they designed their AI.

https://alumni.media.mit.edu/~jorkin/gdc2006_orkin_jeff_fear.pdf

Nothing about the design is fundamentally different from how AI is designed now. Watch the various GDC talks, read Gamasutra articles, and watch Youtube BTS vids on AI design. In fact, most modern day AI does a lot of the same and event takes it a step further




And this isn't taking into account numerous games have various AI systems for things like crowd behavior, animal behavior, etc that all interact with one another in much much open environments

Click to shrink...

fear is just older and came in a time where most other games used the most basic AI so it seemed better but yeah as you say, still using it as a golden standard in 2020 is baffling

 

[Stalker]

AI could fucking ruin you right now the point is games are supposed to be winnable.

 

[Old Gimlet Eyes]

It's interesting argument that people don't want better AI because they would get murdered. Yet Souls series is one of the most popular single player games and Rainbow Six Siege is one of the most popular multiplayer games. In either title you can die really fucking fast, over and over again. Yet they sell like crazy. But people don't want better AI that can kill them faster or in more believable ways? Hmmmmmmm......

 

[Kinthey]

The Last of Us 2 reveal showed how impressive improved AI can feel and how it can effectively be marketed



Unfortunately was also a whole lot of it scripted. To have such an intelligent AI that's able to fully adapt to the player at any moment unfortunately still seems like a pipedream

 

[Starlatine]

It's interesting argument that people don't want better AI because they would get murdered. Yet Souls series is one of the most popular single player games and Rainbow Six Siege is one of the most popular multiplayer games. In either title you can die really fucking fast, over and over again. Yet they sell like crazy. But people don't want better AI that can kill them faster or in more believable ways? Hmmmmmmm......

Click to shrink...

Souls isnt as popular as you think and using Rainbow Six Siege as a counterpoint for AI is just ??????? TTK in multiplayer games is a totally different issues, people love BRs where you can be murdered instantly without barely taking a step but do the same in a SP game and they drop it like a rock

 

[tokkun]

Edit: Also I find it weird that people keep mentioning the power fantasy yet we constantly applaud and hype up extremely difficult online and offline games such as the Souls series. There's a disconnect happening here.

Click to shrink...

The satisfaction people get from Souls games comes from mastering a skill through repetition, learning and building muscle memory. The fact that the AI in Souls games is extremely limited plays into this, because it ensures you will make steady progress if you are willing to invest the time into practicing. It's the same sort of experience as learning to play a new song on a musical instrument.

Online games are a bit different, because the satisfaction there comes from competition instead. However, for a variety of psychological reasons, competing against an AI does not produce the same rush as competing against a human.

 

[Owlet]

i dont think you heard many impressions about the game
cause i sure heard lots of people saying exactly that the Alien was too punishing to be fun

complaints about it being an instakill, complaints about it "cheating" to find you, pick your poison

Click to shrink...

Oh interesting, I never had that experience. Every time it found me it felt like it earned it, but everyone has different experiences.

 

[TheClaw7667]

Enemies that seek cover and run away from grenades and assorted explosives instead of directly into them are by default "harder". F.E.A.R. is indeed slightly more "difficult" than the average shooter thanks to an slightly more intelligent AI. You can't separate one from other. A smarter AI is harder to kill. It's pure logic, i don't know why so many people are arguing against this.

Now if theres anything wrong with that or not is up to anyones taste

Click to shrink...

I was addressing the idea that improvement to AI is something people don't want because it would create an increase in difficulty by making the AI unfair and unfun. The stuff you listed wouldn't do that and is easily changed by difficulty settings. Sure, the AI would be "smarter" if it could take cover instead of just running at the player head-on but not in a way that it would hurt players' enjoyment and not any more difficult.

I also don't agree that seemingly intelligent AI is harder than basic AI. The Last Of Us 2 could have had human AI that just runs at the player but I don't think many people would say that it would create a more enjoyable experience and it would make the game more difficult. Having enemies slowly take cover and take shots at Ellie allows the player time to aim and time for the player to think about their next move. If the enemies just swarmed Ellie the second the player entered a combat scenario the game would be much more difficult and the AI would seem a lot less intelligent.

 

[bushmonkey]

When I think of better AI in games, I don't necessarily mean enemies. It would be fun to have enemies that react more dynamically to situations (while not being as good as the player still) but for me, the main goal should be to make NPCs more intelligent and acting like actual people so you can have better worlds in open-world games that act like infinite sandboxes because the NPCs behave in unexpected ways.